Source: https://saypad.app/dictation-apps-installer-teardown

# We unpacked 8 Mac dictation app installers
On 7 October 2026 we downloaded the current installer of eight Mac dictation apps from their official sites, opened each one read-only and looked at what's inside. Nothing was installed or run. One question mattered most: can the app dictate the moment it's installed, without the internet?

## What's inside each installer

| App | Installer | Unpacked | Biggest parts inside | Speech model inside | Before the first dictation |
|---|---|---|---|---|---|
| Saypad 0.10.0 | 187 MB | 197 MB | Whisper small speech model (190 MB), whisper.cpp | Yes | Nothing – works offline right away |
| Wispr Flow 1.6 | 303 MB | 511 MB | Electron framework (182 MB), app code (127 MB) | No | Sign up; needs internet (cloud) |
| Aqua Voice | 260 MB | 568 MB | Electron framework (372 MB), FluidAudio engine | No | Cloud; the free words come with an account |
| MacWhisper 15.4 | 168 MB | 254 MB | App (92 MB), speaker-separation models (98 MB) | No | Download a model; dictation needs Pro |
| Willow Voice | 85 MB | 118 MB | Interface images (50 MB), app (46 MB) | No | Create an account; cloud by default |
| Superwhisper | 68 MB | 190 MB | ONNX Runtime, MLX and WhisperKit engines | No | Cloud by default, or download a model |
| VoiceInk 2.21 | 51 MB | 110 MB | App (46 MB), whisper.cpp engine | No | Download a model (Parakeet, 494 MB) |
| Spokenly 2.30 | 21 MB | 42 MB | App only | No | Cloud by default, or download a model |

Downloaded from official links and inspected read-only on 7 October 2026. Versions: the latest available that day. “Speech model inside” means a speech-recognition model file; MacWhisper's bundled models are for telling speakers apart, not for recognizing speech.

## Four things we learned

- **Only one installer can dictate on its own.** Saypad ships a light Whisper model (small, 190 MB), so the first dictation works offline. Everyone else needs the cloud or a model download first; Saypad then downloads the larger large-v3-turbo model in the background on Wi-Fi.
- **The biggest installers are browsers.** Wispr Flow and Aqua Voice are built on Electron: the bundled Chromium engine alone is 182 MB and 372 MB unpacked. Their speech recognition happens on servers.
- **Local apps ship engines, not models.** Superwhisper carries three inference engines (ONNX Runtime, Apple's MLX and Argmax's WhisperKit SDK), VoiceInk and MacWhisper carry whisper.cpp, Aqua Voice carries FluidAudio. The model – usually 0.5 to 1.6 GB – comes later, in the setup wizard.
- **Small download, long setup – or the other way round.** Spokenly's 21 MB installer is the lightest, but local dictation needs a separate model download. A one-time 187 MB download that works immediately is a different trade-off, not a better or worse one.

## Which one fits you

- **Dictate right away, offline, for free:** Saypad.
- **AI that rewrites and formats what you say:** Wispr Flow, Aqua Voice or Superwhisper Pro.
- **Open source:** VoiceInk (GPL-3.0).
- **Transcribing files, meetings and podcasts:** MacWhisper.
- **The same app on Windows or your phone:** Superwhisper, Spokenly, Wispr Flow, Aqua Voice or Willow.

For prices and free limits, see [the best free dictation apps for Mac](best-free-dictation-apps-mac.html).

## Questions

**Which Mac dictation app works offline right after install?**
Of the eight we unpacked, only Saypad ships a speech recognition model inside its installer, so the first dictation works without downloading anything. Local-first apps like VoiceInk, Superwhisper and Spokenly can work offline too, but only after you download a model.

**Why is Wispr Flow's installer so big?**
Most of it is Electron – a bundled copy of the Chromium browser engine (182 MB) plus the app's own code (127 MB). Recognition itself runs in the cloud, so there is no speech model inside.

**Does a bigger installer mean better recognition?**
No. The two largest installers, Wispr Flow and Aqua Voice, contain no speech model at all – they are large because of Electron. Recognition quality depends on the model, which most local apps download after install.

**How did you check?**
We downloaded each installer from the vendor's official download link, mounted or unzipped it read-only, and searched for speech model files (.bin, .gguf, .onnx, .mlmodelc, .safetensors) and every file larger than 5 MB. Then we deleted everything. Sizes are as reported by the download server and by du for the unpacked app.

**Are you neutral?**
No – we make Saypad. That's why the method is described above and every other app is credited for what it does better below. Installers change with every release; this is a snapshot of 7 October 2026.